Cantina open-sources apex-flash-1, a GLM-5.3-Flash GRPO post-train for security research: 66.7% pass@1 on 60 held-out vuln tasks at ~1/31 the cost of Opus 5 High

Cantina Security and Yeta Labs released apex-flash-1 (release post, weights), arguing that gating cyber capability in proprietary models mostly constrains defenders, since attackers can already run and modify open models. The model is a GRPO post-train of GLM-5.3-Flash on production-like environments built from real vulnerabilities Cantina found and disclosed: 50 cases, each in guided whitebox, focused whitebox and focused blackbox views (150 tasks), with verifiers checking final target state. Updates used rank-256 LoRA across all experts and routers plus full-parameter updates to 16 activation-selected experts; full updates to all experts collapsed the model.

On Cantina's internal held-out set (60 tasks from 20 unseen cases, first-draw pass@1) it solves 40/60 (66.7%) vs 36/60 for base GLM-5.3-Flash and 43/60 for Claude Opus 5 High, at an estimated $2.38 per run vs $4.56 and $74.68. This is not a public benchmark, and image/video performance is unevaluated. It is designed as a focused worker orchestrated by a larger model, with the Codex harness recommended. Weights are MIT-licensed, with an experimental abliterated variant for researchers. Use only in authorized, isolated environments.